Systems and method for automating detection of regions of machine learning system underperformance

Abstract

In some embodiments, a method includes generating a trained decision tree with a set of nodes based on input data and a partitioning objective, and generating a modified decision tree by recursively passing the input data through the trained decision tree, recursively calculating, for each of the nodes, an associated set of metrics, and recursively defining an association between each of the nodes and the associated set of metrics. A node from a set of nodes of the modified decision tree is identified that violates a user-specified threshold value, associated with a user, for at least one of the metrics. The method also includes causing transmission of a signal to a compute device of the user, the signal including a representation of the identified node.

The invention claimed is: 
     
       1. A method, comprising:
 generating, via a processor, a trained decision tree based on input data and a partitioning objective, the trained decision tree including a plurality of nodes, each node from the plurality of nodes having at least a subset of the input data from a plurality of subsets of the input data; 
 generating, via the processor, a modified decision tree by:
 (a) recursively passing the input data through the trained decision tree to define a recursion, 
 (b) filtering the input data at each split from a plurality of splits of the recursion, 
 (c) recursively calculating, for each node from the plurality of nodes, an associated set of metrics from a plurality of metrics, and 
 (d) recursively defining an association between each node from the plurality of nodes and the associated set of metrics from the plurality of metrics, thereby generating the modified decision tree, at least two of (a), (c), or (d) at least partially overlapping in time, each node from a plurality of nodes of the modified decision tree including a node threshold and metadata that are concurrently stored at each node from the plurality of nodes of the modified decision tree as the trained decision tree is traversed; 
 
 identifying, via the processor, a node from the plurality of nodes of the modified decision tree that violates a user-specified threshold value for at least one metric from the plurality of metrics, the user-specified threshold value associated with a user; and 
 causing transmission of a signal including a representation of the identified node from the plurality of nodes of the modified decision tree from the processor to a compute device of the user. 
 
     
     
       2. The method of  claim 1 , wherein the signal further includes a representation of the set of metrics from the plurality of metrics associated with the identified node from the plurality of nodes of the modified decision tree. 
     
     
       3. The method of  claim 1 , further comprising not further attempting to identify violations of the user-specified threshold value in downstream nodes from the plurality of nodes of the modified decision tree, in response to identifying the node that violates the user-specified threshold value. 
     
     
       4. The method of  claim 1 , wherein each set of metrics from the plurality of metrics includes a representation of at least one of: an impurity, a number of samples, an entropy, split information, a feature rule set of an associated node path, an accuracy, a precision, a recall, or an F1 score. 
     
     
       5. The method of  claim 1 , wherein the input data is streaming data and the generating the trained decision tree is performed over a predefined time period. 
     
     
       6. The method of  claim 1 , wherein the input data is streaming data associated with a predefined time period. 
     
     
       7. The method of  claim 1 , wherein the input data includes a plurality of classification predictions, and the generating the trained decision tree includes labelling each subset of the input data from the plurality of subsets of the input data based on an accuracy of a subset of classification predictions, from the plurality of classification predictions, associated with that subset of the input data. 
     
     
       8. The method of  claim 1 , wherein the input data includes a plurality of regression outputs, and the generating the trained decision tree includes labelling each subset of the input data from the plurality of subsets of the input data based on a prediction error of a subset of regression outputs, from the plurality of regression outputs, associated with that subset of the input data. 
     
     
       9. The method of  claim 1 , wherein the input data includes the metadata and the partitioning objective is defined, at least in part, by the metadata. 
     
     
       10. The method of  claim 1 , wherein the generating the trained decision tree is further based on an information criterion. 
     
     
       11. The method of  claim 10 , wherein the information criterion is a Gini impurity. 
     
     
       12. The method of  claim 1 , wherein the recursively passing the input data through the trained decision tree is performed for each ground truth from a plurality of ground truths. 
     
     
       13. The method of  claim 1 , wherein the input data includes tabular data. 
     
     
       14. The method of  claim 1 , wherein the input data includes low-dimensional representations of images. 
     
     
       15. The method of  claim 14 , wherein the low-dimensional representations of the images are generated using at least one of a convolutional neural network or transfer learning. 
     
     
       16. The method of  claim 1 , wherein the input data includes low-dimensional representations of text for natural language processing. 
     
     
       17. The method of  claim 16 , wherein the low-dimensional representations of the text are generated using at least one of a language model or transfer learning. 
     
     
       18. The method of  claim 1 , wherein each node from the plurality of nodes of the modified decision tree represents a collection of datapoints that are filtered by accumulated ranges of features of the input data along a path from a root node to that node. 
     
     
       19. The method of  claim 1 , wherein the identified node is an intermediate node of the modified decision tree. 
     
     
       20. The method of  claim 1 , wherein the identified node is positioned between a root node of the modified decision tree and a terminal leaf node of the modified decision tree. 
     
     
       21. A method, comprising:
 generating, via a processor, a modified decision tree by:
 (a) recursively passing, via the processor, input data through a trained decision tree to define a recursion, 
 (b) filtering the input data at each split from a plurality of splits of the recursion, recursively calculating, via the processor and for each node from a plurality of nodes of the trained decision tree, an associated set of metrics from a plurality of metrics, and 
 (d) recursively defining, via the processor, an association between each node from the plurality of nodes and the associated set of metrics from the plurality of metrics, thereby generating the modified decision tree, at least two of (a), (c) or (d) at least partially overlapping in time, each node from a plurality of nodes of the modified decision tree including a node threshold and metadata that are concurrently stored at each node from the plurality of nodes of the modified decision tree as the trained decision tree is traversed; 
 
 identifying, via the processor, a node from the plurality of nodes of the modified decision tree that violates a user-specified threshold value for at least one metric from the plurality of metrics, the user-specified threshold value associated with a user; and 
 causing transmission of a signal including a representation of the identified node from the plurality of nodes of the modified decision tree from the processor to a compute device of the user. 
 
     
     
       22. A method, comprising:
 training, via a processor, a decision tree based on input data and a partitioning objective, to produce a trained decision tree that includes a plurality of nodes, each node from the plurality of nodes including a portion of the input data; 
 generating, via the processor, a modified decision tree by:
 (a) recursively passing, via the processor, the input data through the trained decision tree to define a recursion, 
 (b) filtering the input data at each split from a plurality of splits of the recursion, 
 (c) recursively calculating, via the processor and for each node from the plurality of nodes, an associated set of metrics from a plurality of metrics, and 
 (d) recursively defining, via the processor, an association between each node from the plurality of nodes and the associated set of metrics from the plurality of metrics, thereby generating the modified decision tree, at least two of (a), (c) or (d) at least partially overlapping in time, each node from a plurality of nodes of the modified decision tree including a node threshold and metadata that are concurrently stored at each node from the plurality of nodes of the modified decision tree as the trained decision tree is traversed; 
 
 identifying, via the processor, a node from the plurality of nodes of the modified decision tree that violates a user-specified threshold value for at least one metric from the plurality of metrics, the user-specified threshold value associated with a user; and 
 causing transmission of a signal including a representation of the identified node from the plurality of nodes of the modified decision tree from the processor to a compute device of the user. 
 
     
     
       23. The method of  claim 22 , wherein the input data includes data generated using transfer learning.